Scale the bus speed within the peripheral instead of on the system bus
as scaling on the system bus slows down general register access to the
peripheral.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Simultaneously define the hardware pin and the timer registers. This
makes the table a little easier to understand.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Increase the amount of time to "busy wait" in the timer dispatch loop
to 2 micro-seconds. This, in practice, causes stepper step and unstep
events to occur within a single hardware interrupt. Doing that helps
stabilize the single active stepper performance benchmarks.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
The atmega328p is basically the same as the atmega328 - add explicit
support for it so that avrdude doesn't complain while flashing.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Sometime bossac reports an error during chip restart. It appears this
error is spurious - just suppress it.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Scaling the timer on the system bus has a notable impact on
performance. So, scale the timer within the timer hardware itself.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
It appears the sam3x8e has a primitive cache prefetch mechanism (it
prefetches 32 bytes at a time aligned to a 16 byte boundary).
Aligning the main loop in timer_dispatch_many() to a 16 byte boundary
significantly improves performance.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
If sched_try_shutdown() is not inlined into timer_dispatch_many() it
can have a significant impact on benchmarks. (Even though the
function isn't called in practice, the call causes gcc to organize the
function differently.)
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Some older versions of gcc need the __always_inline directive in order
to inline timer_read_time. Inlining that function is important for
performance on the stm32f1.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
As long as timer_read_time() is inlined, I get better performance with
gcc -O2 optimization. The binary is also dramatically smaller and O2
better matches the other platforms.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
It's possible for the code to stay in the timer irq for up to 100ms,
so serial irqs should have a higher irq priority to prevent them from
being starved. (The timer code disables irqs during event dispatch,
so serial irqs would only be permitted between events anyway.)
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
The usb_cdc code will arrange to call usb_read_ep0_setup() after a
shutdown, so no need to have an explicit shutdown handler.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Call usb_read_bulk_out() before calling command_find_and_dispatch() as
this optimizes the common case where each usb packet contains a single
new message block.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Recent code additions have increased the stack usage and it is no
longer clear that everything will fit in 128 bytes. Increase all
targets to 256 bytes (the atmega2560 was already at 256 bytes).
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Fix travis-ci build failure by forcing the data field of the struct
usb_string_descriptor to be the same type as a 16bit unicode string.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Yannic Schroeder